Lco road to 8000 or 10000! self play elo.

[Nay Lin Tun]

The training of convolutional neural networks (CNN) is "Challenge yourself to beat Yesterday", and 30xxx network is doing her own job amazingly, currently 7200+ self elo and ongoing..
https://ibb.co/fpiVfV

P.S, Testers say 30xx network is still -100 to -150 elo beyond 11248.

[whereagles]

Leela has a promising career in politics if chess fails

[alex67a]

Mah!
I played a mini-match Lc0 Network 7200 elo Vs Lco Network N. 11149 elo 4851...
Four games, 30 sec to move
Lc0 with old NN is stronger

[mar]

Four games, 30 sec to move
Lc0 with old NN is stronger

This is hilarious, play 1 game next time.
I wish I knew that it's sufficient to test with just a handful of games, that would've saved me tremendous amount of time.

[Milos]

Four games, 30 sec to move
Lc0 with old NN is stronger

This is hilarious, play 1 game next time.
I wish I knew that it's sufficient to test with just a handful of games, that would've saved me tremendous amount of time.

The only hilarious thing is Lc0 self-testing.

[Laskos]

Four games, 30 sec to move
Lc0 with old NN is stronger

This is hilarious, play 1 game next time.
I wish I knew that it's sufficient to test with just a handful of games, that would've saved me tremendous amount of time.

To check the truthfulness of 2300+ Elo difference, even 3 games can be more than enough for high confidence of its falsification.

[alex67a]

This is hilarious, play 1 game next time.
I wish I knew that it's sufficient to test with just a handful of games, that would've saved me tremendous amount of time.

Obviously it is a simple test, I do not pretend it is super precise, it is only indicative
I do not have time to play lots of games
But Lc0 plays against himself doing similar results (+3 = 1 -0)
75% of the games win, if they were tied, I would agree to say that we need to play a lot more games
But with these results I think the conclusions are clear
However I did do several mini-matches of 4 games with the latest NN very high and different from each other, but the old NN wins every time

[MikeB]

P.S, Testers say 30xx network is still -100 to -150 elo beyond 11248.

Fake news....

anyone can verify or disprove your claims after even just a few games, I believe it's actually worse...

Code: Select all

Rank Name             Rating   Δ     +    -     #     Σ    Σ%     W    L    D   W%    =%   OppR 
---------------------------------------------------------------------------------------------------------
   1 Lc0 v0.19.0 11261   3198   0.0   84   84    50   38.5  77.0   33    6   11  66.0  22.0  3002 
   2 Lc0 v0.19.0 31493   3002 195.7   84   84    50   11.5  23.0    6   33   11  12.0  22.0  3198
---------------------------------------------------------------------------------------------------------

custom openings that may exaggerate Elo differences ( due to the unbalance nature of the openings )..

[chrisw]

I hope this isn't happening ....

Code: Select all

import random

program_new_trained_version_elo = 0
program_old_version_elo = 0
current_best_leela_elo = 0
we_didnt_get_bored_yet = 0
while (we_didnt_get_bored_yet < 2500):
    # play a match, the programs are the same because learning stalled, but the results are always a bit random ....
    random_result = (random.randint(0,1))
    
    if (random_result == 1):
        # if new trained version happens to be randomly better, like 50% of the time ...
        # keep new trained version as leela, and increase the elo a bit (because Leela won this one)
        program_new_trained_version_elo += 5
        current_best_leela_elo = program_new_trained_version_elo
    # else:
        # new trained version happens to not be randomly better
        # don't update leela    
    we_didnt_get_bored_yet += 1
# we got bored ....    
print('leela self_play elo has liftoff, Houston! Elo rating=' + str(current_best_leela_elo))

run program ....
leela self_play elo has liftoff, Houston! Elo rating=6250

[alex67a]

Fake news....

anyone can verify or disprove your claims after even just a few games, I believe it's actually worse...

Code: Select all

Rank Name             Rating   Δ     +    -     #     Σ    Σ%     W    L    D   W%    =%   OppR 
---------------------------------------------------------------------------------------------------------
   1 Lc0 v0.19.0 11261   3198   0.0   84   84    50   38.5  77.0   33    6   11  66.0  22.0  3002 
   2 Lc0 v0.19.0 31493   3002 195.7   84   84    50   11.5  23.0    6   33   11  12.0  22.0  3198
---------------------------------------------------------------------------------------------------------

custom openings that may exaggerate Elo differences ( due to the unbalance nature of the openings )..

Yes

I used NN 11149, I think it is slightly higher than 11261 and 11248